THE KNOCK-KNOCK CARBON COST

The idea that AI lives "in the cloud" is one of the most effective pieces of misdirection in modern technology. The cloud is a metaphor. What it describes is a network of massive, physical, power-hungry buildings located in real places, consuming real resources, generating real emissions.

Every AI query you make activates physical machinery. Servers draw electricity. Cooling systems consume water. Heat is expelled. The experience on your screen is frictionless. The infrastructure making it possible is anything but.

This episode maps that hidden footprint — from the scale of energy consumption to the local communities living next to the facilities that power the AI you use every day.

This episode moves through five connected areas of investigation. Each one builds on the last — from understanding what the infrastructure actually is, to measuring its scale, to the local communities absorbing its cost, to why the full picture is so hard to see, and finally to the collision between AI's growth trajectory and the climate commitments it's undermining.

The argument isn't that AI is inherently bad. It's that the current pace of expansion is outrunning the infrastructure needed to power it sustainably — and that the people making decisions about that expansion have strong incentives to keep the costs invisible.

A data center is a temperature-controlled building housing rows of computer servers, data storage drives, and networking equipment — plus the power and cooling systems that keep them running. They range from the size of a large office building to the size of multiple city blocks.

Amazon alone has more than 100 data centers worldwide, each with roughly 50,000 servers. Google, Microsoft, and Meta operate comparable footprints. The infrastructure powering the AI you use daily is distributed across hundreds of these facilities — all consuming electricity, all requiring water for cooling, all generating heat that must be expelled.

This is the origin of the episode title. Every AI interaction — the "knock knock" of your prompt — activates physical machinery somewhere. The machine answers. Resources are consumed. The conversation continues. Multiply that by billions of daily queries and the aggregate cost becomes significant.

Researchers at UC Riverside estimated that each 100-word AI prompt uses roughly one bottle of water — about 519 milliliters — for cooling. That figure accounts for the water evaporated by cooling systems during the computation required to generate a response. With billions of users worldwide, the cumulative water demand is enormous.

The defining philosophy of modern AI development has been scaling — the observation that larger models with more parameters, trained on more data with more compute, tend to perform better. This drove an extraordinary expansion in model size between 2019 and 2024.

GPT-3 had 175 billion parameters when it launched in 2020. Models since have grown dramatically beyond that. Each order-of-magnitude increase in model scale brings a corresponding increase in the compute required to train it — and with that compute comes energy, water, and emissions.

A standard web search retrieves a result that already exists — an index lookup returning cached pages. It is computationally cheap. A generative AI query does something fundamentally different: it generates a new response, token by token, requiring intensive computation at every step.

Estimates suggest a single generative AI query consumes roughly 10 times the energy of a standard search query. As AI becomes the default interface for information retrieval — replacing search in many contexts — the aggregate energy cost of everyday information lookup rises dramatically.

BLOOM — a 176-billion parameter language model developed by Hugging Face and the BigScience research initiative — is one of the most carefully documented AI models in terms of its environmental footprint. Researchers published a peer-reviewed paper in 2022 calculating its full lifecycle carbon cost.

Training BLOOM emitted approximately 24.7 tonnes of CO2 equivalent — roughly comparable to driving a car around the planet several times. When the full lifecycle is included (hardware manufacturing, deployment, ongoing inference), the figure roughly doubles to 50.5 tonnes.

BLOOM is notable precisely because it was trained on a French nuclear-powered supercomputer — one of the lower-carbon options available. Models trained on fossil-fuel-heavy grids emit significantly more.

The comparison between BLOOM and GPT-3 is instructive: both have similar parameter counts, but GPT-3 emitted roughly 20 times more CO2 during training. The difference isn't primarily model architecture — it's where the training happened and what powered it.

BLOOM was trained on a French supercomputer powered largely by nuclear energy. GPT-3 was trained on less efficient hardware connected to a higher-carbon grid. This means the location of a data center — and the energy mix of the local grid — can be as significant an environmental factor as the model's size.

The environmental impacts of data centers are not abstract global statistics. They are experienced locally — by the people who live near the facilities, rely on the same water utilities, and share the same power grid.

Northern Virginia hosts the densest concentration of data centers anywhere in the world — roughly 300 facilities in a handful of counties. In Loudoun County alone, data center tax revenues approach $900 million annually — but residents and water officials are increasingly raising concerns about the strain on local resources that accompanies that revenue.

Cooling systems — the banks of fans, chillers, and air handling units that prevent servers from overheating — operate continuously. They are not intermittent. They do not observe quiet hours. For communities adjacent to large data centers, this translates to a permanent mechanical soundscape that residents have described as impossible to escape.

This is not speculative. Residents near data center clusters in Northern Virginia, The Dalles (Oregon), and elsewhere have filed complaints and initiated legal proceedings related to noise, water use, and loss of access to local resources. The Lincoln Institute has documented these cases as part of broader research on the community costs of data center expansion.

A Meta data center in Newton County, Georgia uses 500,000 gallons of water per day — 10% of the entire county's water consumption. New permits for the same area could bring usage up to 6 million gallons per day, more than doubling what the entire county currently consumes. These are not projections — they are documented permit applications.

At the global scale, ScienceDirect researchers estimated AI systems' water footprint could reach 312 to 765 billion litres in 2025 — a range equivalent to the world's annual bottled water consumption. Training GPT-3 alone evaporated roughly 700,000 litres of clean freshwater in Microsoft's US data centers.

The pace at which data centers are being built has outrun the capacity of renewable energy to power them. When a utility company cannot meet new demand from clean sources, it falls back on existing fossil fuel plants — or delays retiring them. In some cases, it has meant recommissioning plants that were scheduled to close.

MIT researcher Noman Bashir has stated plainly: "The demand for new data centers cannot be met in a sustainable way." The electricity grid in many regions was designed around residential and commercial demand profiles that look nothing like the constant, massive load that hyperscale data centers require.

Researchers and journalists attempting to quantify the environmental impact of AI face a consistent obstacle: the companies operating the infrastructure don't disclose what they'd need to disclose for accurate measurement. Water usage, energy consumption by workload type, and expansion plans are routinely withheld, redacted, or simply never reported.

The Lincoln Institute documented that data centers "are coming in with non-disclosure agreements" in communities across the US — meaning local officials who approve permits often do not know, and cannot share, the full resource footprint of the facilities they're approving.

The script is careful here — and this episode is too. There is a meaningful distinction between what is documented and what is alleged. The documented behaviors include: records being withheld or redacted, data classified as trade secrets, land purchased through shell entities ahead of community notification, and NDAs in permitting processes. These are reported facts.

Whether these behaviors reflect a deliberate strategy to conceal environmental impact is interpretation — not established fact. Companies have legal and competitive reasons to protect operational data that have nothing to do with intent to deceive. Both things can be true simultaneously.

This episode deliberately distinguishes between what sources directly support and what moves into allegation. The distinction isn't caution for its own sake — it's because the documented record is already damaging enough that adding unsupported allegations actually weakens the argument.

AI systems demonstrably consume enormous amounts of energy and water. Data centers demonstrably strain local communities. Environmental disclosures are demonstrably inadequate. These facts don't require allegations about motive to be serious — they're serious on their own terms.

The AI industry and the climate commitments of the same companies operating it are moving in opposite directions. The logic of AI development says: build bigger, build faster, train more, deploy everywhere. The logic of climate responsibility says: measure everything, reduce consumption, eliminate fossil dependence.

Cornell's 2025 research team put it plainly: "The AI infrastructure choices we make this decade will decide whether AI accelerates climate progress or becomes a new environmental burden." That window is not theoretical — it is the period during which data center locations, cooling technologies, and energy contracts are being locked in for the next 20 years.

The question the episode ends on is not rhetorical. It is a genuine open problem that regulators, engineers, and citizens will have to answer — or have answered for them by default.